Harald

A table of answers, each one with its quote.

One row per contract, one column per question. Behind every green cell is a passage of the contract that code found, word for word.

The Harald review table: contracts in rows, questions in columns, each cell marked verified, needs review or not found

The review table on the demo set, which is made of contracts from the public CUAD set.

A cell has three states, and code sets them

The model proposes an answer and a quote. What the cell is allowed to say about itself is decided afterwards, by a rule a person can read.

Verified

The quote is one continuous passage of the contract, with the same letters, digits and symbols. Capitals and spacing may differ. A quote joined with an ellipsis is refused.

Needs review

The model answered, and code did not find the quote, or the short answer says something the quote does not. A person decides.

Not found

The model found no clause that answers the question. The cell says so and claims nothing.

Ask a question across the whole set

Ask Harald answers from the verified cells of the run and from nothing else. It uses no language model, so it cannot add a sentence the table does not hold. Each line of the answer names its contract.

Ask Harald: a question typed above the table and an answer that lists contracts with their verified quotes

What the reviewer does

The quote opens beside the contract, at its place in the text. Every decision is recorded with the person and the time.

  • Accept or reject an answer
  • Run one cell again
  • Mark a cell as reviewed; a padlock shows who and when
  • Filter the table to the cells that need review
  • Test the check on a quote of your own
  • Export the table as CSV

The ten questions Harald asks today

Governing law, change of control, anti-assignment, termination for convenience, non-compete, exclusivity, cap on liability, renewal term, notice to stop renewal, audit rights.

The files it reads

Word (.docx and old .doc), PDF and text. A scanned page is recognised on the computer itself, in English and Italian, up to 100 pages a document, and the contract is marked as recognised from a scan.

Limit. The check compares words, not meaning. A quote can be real and still not answer the question, which is why precision is lower than recall. See the figures

Break a quote yourself.

The check on the first page runs in your browser. Add one word to the quote and watch the cell lose its green.